Expert Systems With Applications 86 (2017) 64–76

Contents lists available at ScienceDirect

Expert Systems With Applications

journal homepage: www.elsevier.com/locate/eswa

Multilevel thresholding using grey wolf optimizer for image

segmentation
Abdul Kayom Md Khairuzzaman, Saurabh Chaudhury∗
a
Department of Electrical Engineering, National Institute of Technology (NIT) Silchar, Silchar 788010, Assam, India

a r t i c l e i n f o a b s t r a c t

Article history: Multilevel thresholding is one of the most important areas in the field of image segmentation. How-
Received 14 March 2016 ever, the computational complexity of multilevel thresholding increases exponentially with the increas-
Revised 14 April 2017
ing number of thresholds. To overcome this drawback, a new approach of multilevel thresholding based
Accepted 15 April 2017
on Grey Wolf Optimizer (GWO) is proposed in this paper. GWO is inspired from the social and hunting
Available online 17 April 2017
behaviour of the grey wolves. This metaheuristic algorithm is applied to multilevel thresholding prob-
Keywords: lem using Kapur’s entropy and Otsu’s between class variance functions. The proposed method is tested on
Multilevel thresholding a set of standard test images. The performances of the proposed method are then compared with im-
Image segmentation proved versions of PSO (Particle Swarm Optimization) and BFO (Bacterial Foraging Optimization) based
Grey wolf optimizer multilevel thresholding methods. The quality of the segmented images is computed using Mean Struc-
Kapur’s entropy tural SIMilarity (MSSIM) index. Experimental results suggest that the proposed method is more stable
Otsu’s threshold
and yields solutions of higher quality than PSO and BFO based methods. Moreover, the proposed method
is found to be faster than BFO but slower than the PSO based method.
© 2017 Published by Elsevier Ltd.

1. Introduction
Segmentation partitions an image into some meaningful regions
or extracts its boundaries. It is often the preprocessing stage of
higher level processing such as, image analysis, and object recog-
nition and computer vision. Therefore, the performance of higher
level processing system depends on the accuracy of segmentation
technique used.
Histogram based thresholding is a simple and the most pop-
ular image segmentation technique. Some survey on threshold-
ing techniques can be found in (Mehmet Sezgin, 2004; Pal &
Pal, 1993; Sahoo & Soltani, 1988). Thresholding techniques can be
broadly classified into local and global. Global threshold selection
techniques are most popular due to their simplicity and effec-
tiveness. Among the global thresholding techniques, Kapur, Sahoo
and Wong (1985) and Otsu (1979) methods are the most popular
ones. Otsu method maximizes the between class variance function
whereas, Kapur method maximizes posterior entropy of the seg-
mented classes to find optimum thresholds. Computational com-
plexity of Kapur and Otsu methods increases exponentially with
the increasing number of thresholds due to exhaustive search used
for finding the thresholds.
∗
Corresponding author.
E-mail addresses: kayomabdul@gmail.com (A.K.M. Khairuzzaman),
saurabh@nits.ac.in (S. Chaudhury).
Liao, Chen and Chung (2001) proposed a fast recursive al-
gorithm with a look up table for multilevel Otsu thresholding.
Yin and Chen (1997) proposed an iterative scheme for multilevel
thresholding. Reddi, Rudin, and Keshavan (1984) used a criterion
function which was derived by assuming grey level histogram of
an image to be a continuous probability function. Although these
methods reduce the computational cost to some extent, processing
time is still an issue. A statistical recursive algorithm proposed by
Arora, Acharya, Verma, and Panigrahi (2008) uses mean and vari-
ance in order to perform multilevel image segmentation. It starts
from the two extreme ends of the histogram and recursively ap-
ply the algorithm until there is no significant improvement in the
segmented image. Although this method is very fast (<0.3 s), it as-
sumes the image histogram as a Gaussian distribution and uses
only the even number of thresholds.
Nature has always been an inspiration in solving computation-
ally hard problems. Genetic Algorithm (GA) is applied in multi-
level thresholding (Yin, 1999) using Otsu and Kapur functions. This
method gives a very good segmentation with substantial reduction
in the computational cost of multilevel thresholding. Differential
Evolution (DE) along with Gaussian approximation function is ap-
plied in multilevel thresholding (Cuevas, Zaldivar & Pérez-cisneros,
2010). A synergetic DE algorithm is applied in multilevel thresh-
olding (Ali, Ahn & Pant, 2014). This method outperforms Particle
Swarm Optimization (PSO), GA, and DE based methods. The com-
putational cost is greatly reduced but still it is high (∼2 s). A hybrid

http://dx.doi.org/10.1016/j.eswa.2017.04.029
0957-4174/© 2017 Published by Elsevier Ltd.
 A.K.M. Khairuzzaman, S. Chaudhury / Expert Systems With Applications 86 (2017) 64–76 65

DE algorithm using Otsu function is used for multilevel threshold- ment an image. However, selecting the appropriate values of these
ing (Brest, Mlakar & Poto, 2016). This method is very fast (∼0.5 s). thresholds is very important in obtaining a good segmentation. Op-
Maitra and Chatterjee has applied a comprehensive learning based timal threshold selection techniques search for thresholds by op-
hybrid PSO algorithm in multilevel thresholding (Maitra & Chat- timizing (either minimizing or maximizing) an objective function.
terjee, 2008a). This algorithm takes care of stagnation problem of Kapur’s entropy and Otsu’s between class variance based methods
PSO algorithm. Gao, Xu, Sun, and Tang (2010) has proposed a quan- are the two most widely used optimal thresholding techniques. In
tum behaved PSO algorithm for multilevel thresholding using Otsu the following subsections we give a brief formulation of the above
function. Although this method has greatly reduced the compu- two multilevel thresholding techniques.
tational complexity of multilevel thresholding, the computational Let us assume that there are L numbers of grey levels in a given
time is still high (>5 s). Artificial bee colony algorithm is used to image which are in the range{0, 1, 2, ..., (L − 1 )}. Let N be the to-
optimize a Gaussian mixture model for multilevel image segmen- tal number of pixels in the image. A particular gray level ioccurs
tation (Cuevas, Sencion, Zaldivar, Perez-Cisneros & Sossa, 2012). An ni times in the image. The gray level histogram of the image can
improved version of bat algorithm is applied in multilevel thresh- be normalized and regarded as a probability distribution. We de-
olding (Alihodzic & Tuba, 2014). A cuckoo search algorithm is also fine the probability of occurrence of a grey levelias, pi = ni /N. The
applied in multilevel thresholding using Kapur’s entropy as crite- following subsections describe Kapur’s entropy and Otsu’s between
rion function to segment satellite images (Bhandari, Singh, Kumar class variance functions in brief.
& Singh, 2014). Sathya and Kayalvizhi have applied Bacterial Forag-
ing Optimization (BFO) algorithm and its modified version (Sathya 2.1. Otsu (between class variance) method
& Kayalvizhi, 2011b; 2011c) in multilevel thresholding using Otsu
and Kapur functions. These two methods also substantially reduce Otsu method (Otsu, 1979) is a nonparametric and unsuper-
the computational complexity of multilevel thresholding. But com- vised automatic threshold selection technique. It selects optimum
putational time is more than 2 s for Otsu function, whereas, it is thresholds by maximizing between class variance of the segmented
more than 6 s for Kapur’s function. Success of these nature in- classes. Let there be m number of thresholds [t1 , t2 , ..., tm ] to be se-
spired algorithms in multilevel thresholding applications and the lected. These thresholds subdivide the image intom + 1classes: C0 ,
quest for further reduction in computational complexity of multi- C1 , C2 ,…, Cm by maximizing the objective function,
level thresholding have inspired the authors to investigate the mer-
J1 (t1 , t2 , ..., tm ) = σ02 + σ12 + σ22 + ... + σm2 (1)
its of Grey Wolf Optimizer (GWO) algorithm in multilevel thresh-
olding. where
Multilevel thresholding has been applied in various image seg- 1 −1
t 1 −1
t
mentation applications. For example, Maitra and Chatterjee has σ02 = ω0 (μ0 − μT )2 , ω0 = p i , μ0 = i pi /ω0
proposed a magnetic resonance image segmentation method us- i=0 i=0
ing multilevel thresholding (Maitra & Chatterjee, 2008b). This
2 −1
t 2 −1
t
method has applied BFO algorithm using Kapur’s entropy func-
σ12 = ω1 (μ1 − μT )2 , ω1 = p i , μ1 = i pi /ω1
tion for multilevel thresholding. Similarly, Sathya and Kayalvizhi
i=t1 i=t1
(2011a) has presented a magnetic resonance image segmentation
method based on BFO algorithm. Bhandari et al. (2014) has applied 3 −1
t 3 −1
t
multilevel thresholding using Cuckoo Search for segmenting satel- σ22 = ω2 (μ2 − μT )2 , ω2 = p i , μ2 = i pi /ω2
lite images. Sarkar, Das, and Sinha (2016) has proposed a multi- i=t2 i=t2
level thresholding based satellite hyper-spectral image segmenta-
L−1
 L−1

tion technique using Renyi’s entropy aided with differential evolu-
σm2 = ωm (μm − μT )2 , ωm = p i , μm = i pi /ωm
tion.
i=tm i=tm
This paper presents a new method of multilevel thresholding
based on Grey Wolf Optimizer (GWO). The proposed method se- σ02 , σ12 , σ22 ,…, σm2 are the variances. ω0 , ω1 , ω2 …,ωm are the class
lects the optimal set of thresholds using either Otsu’s between probabilities. μ0 , μ1 , μ2 ,…, μm are the mean levels of the seg-
class variance or Kapur’s entropy function. The main contributions mented classes: C0 , C1 , C2 ,…, Cm respectively. If μT is the mean
of this paper are: (1) the application of GWO for optimal multilevel intensity for the whole image, then we have, ω0 μ0 + ω1 μ1 +
thresholding using Otsu and Kapur methods. The result of experi- ω2 μ2 + ... + ωm μm = μT and ω0 + ω1 + ω2 + ... + ωm = 1.
mentation suggest that GWO gives better result compared to BFO Eq. (1) is maximized to find the optimum thresholds t1, t2, …,tm.
and PSO based methods, and (2) the computational complexity of Since the GWO algorithm minimizes an objective function, there-
multilevel thresholding is greatly reduced. fore, the objective function in Eq. (1) is converted to an equivalent
The remainder of the article is organized as follows. minimization function by taking the inverse of it. The correspond-
Section 2 describes the multilevel thresholding problem. It ing minimization function is taken as,
also describes Kapur’s entropy and Otsu’s between class variance
J1 (t1 , t2 , ..., tm ) = 1/(1 + J1 (t1 , t2 , ..., tm )) (2)
functions for multilevel thresholding. Section 3 gives an overview
of GWO followed by its mathematical model. Section 4 describes Minimizing the function in Eq. (2) is equivalent to maximizing
the proposed GWO based multilevel thresholding method followed the between class variance function in Eq. (1). After finding the op-
by its pseudo code. Section 5 describes the experimental environ- timum solution, the objective function value is converted back into
ment of the proposed method. Results and discussion are given in the maximization form as in Eq. (1) and presented in the paper.
Section 6. Finally, the conclusions are provided in Section 7.
2.2. Kapur’s (entropy criterion) method
2. Multilevel thresholding
Kapur’s method (Kapur et al., 1985) selects the optimum
Bi-level thresholding partitions an image into two classes: the thresholds by maximizing the entropy of the segmented classes. It
object and the background. If the image is complex and contains uses Shannon’s concept of entropy. Shannon’s function (Shannon,
multiple objects then bi-level thresholding is not very effective. 1948) says that the information content of an event is inversely
In such a situation, multilevel thresholding is often used to seg- proportional to its probability of occurrence. Kapur et al. define
66 A.K.M. Khairuzzaman, S. Chaudhury / Expert Systems With Applications 86 (2017) 64–76

the entropy of an image assuming that an image is entirely repre-
sented by its grey level histogram. If there are m number of thresh-
olds [t1 , t2 , ..., tm ] to be selected which subdivide the image into
the classes: C0 , C1 , C2 ,…, Cm , then Kapur’s method does it by max-
imizing the objective function,
3.2.1. Social hierarchy
To model the social hierarchy of the grey wolves, the fittest so-
lution is considered as the α . The next two best solutions are con-
sidered as the β and the δ . The rest are ω wolves. The optimization
process is guided by α , β , and the δ . The ω wolves follow them.

J2 (t1 , t2 , ..., tm ) = H0 + H1 + H2 + ... + Hm (3) 3.2.2. Encircling of prey

Grey wolves encircle the prey during hunting. The encircling
where, behaviour is modelled as,
−
→ −
→ −
→
1 −1
t 1 −1
t
D = |C · X p (t ) − X (t )| (5)
H0 = − ( pi /ω0 ) ln( pi /ω0 ), ω0 = pi
i=0 i=0 −
→ −
→ −
→
X (t + 1 ) = X p (t ) − A · D (6)
2 −1
t 2 −1
t
−
→
H1 = − ( pi /ω1 ) ln( pi /ω1 ), ω1 = pi where t is the current iteration, A and C are the coefficients. X p
i=t1 i=t1 −
→
and X are the position vectors of the prey and a grey wolf respec-
3 −1
t 3 −1
t tively. The values of the coefficients A and C are calculated as,
H2 = − ( pi /ω2 ) ln( pi /ω2 ), ω2 = pi
A = 2a · r1 − a (7)
i=t2 i=t2
L−1
 L−1

Hm = − ( pi /ωm ) ln( pi /ωm ), ωm = pi C = 2 · r2 (8)
i=tm i=tm The value of a is linearly decreased from 2 to 0 during the
iterations, while, r1 and r2 are random numbers in the range
H0 , H1 , H2 ,…, Hm are the Kapur’s entropies. ω0 , ω1 , ω2 ,…, ωm
[0, 1]. Grey wolves update their positions around the prey using
are the class probabilities of the segmented classes: C0 , C1 , C2 ,…,
Eqs. (5) and (6). In m-dimensional search space the grey wolves
Cm respectively.
move in hyper-spheres around the best solution obtained so far.
Eq. (3) is maximized to find the optimum thresholds t1, t2, …,tm .
Similar to Eq. (1), Eq. (3) is also converted to an equivalent min- 3.2.3. Hunting
imization function by taking the inverse of it. The corresponding The hunting process is guided by the alpha and sometimes also
minimization function is taken as, by beta, and delta. It is assumed that the α , β , and the δ have bet-
ter knowledge about the location of the prey (optimum solution).
J2 (t1 , t2 , ..., tm ) = 1/(1 + J2 (t1 , t2 , ..., tm )) (4)
Therefore, the other wolves update their positions according to the
Minimization of the function in Eq. (4) corresponds to the max- positions of the alpha, beta, and the delta. The mathematical model
imization of Kapur’s entropy function in Eq. (3). After finding the for hunting is given as,
optimum solution, the objective function value is converted back −
→ −
→ → −
− → −
→ → −
− → −
→ −
→
D α = |C1 · X α − X |, D β = |C2 · X β − X |, D δ = |C3 · X δ − X |,
into the maximization form as in Eq. (3) and presented in the pa-
per. (9)

−
→ −
→ → −
− → −
→ → −
− → −
→ −
→
3. Grey wolf optimizer X 1 = X α − A1 · D α X 2 = X β − A2 · D β X 3 = X δ − A3 · D δ
(10)
Grey wolf optimizer (GWO) is a metaheuristic proposed by −
→ − → −
→
Mirjalili, Mohammad, and Lewis (2014). A brief overview and the where X α , X β , and X δ are the position vectors of alpha, beta,
mathematical model of GWO are given in the following subsec- and the delta respectively. A1 , A2 , A3 , C1 , C2 , and C3 are the co-
tions. efficients calculated using Eqs. (7) and (8) with a different set of
random numbers.
The position of a wolf is updated as,
3.1. Overview −
→ −
→ −
→ −
→
X (t + 1 ) = ( X 1 + X 2 + X 3 )/3 (11)
Grey wolf optimizer is inspired by the social hierarchy and the Eq. (11) shows how a grey wolf (search agent) updates its posi-
hunting technique of grey wolves. Grey wolves live in a highly or- tion according to alpha, beta, and delta in a search space.
ganized pack. The average pack size ranges from 5 to 12. Normally,
there are four different ranks of wolves in a pack. These are al- 3.2.4. Attacking the prey (exploitation)
pha (α ), beta (β ), delta (δ ), and omega (ω) wolves. The alphas (a When the prey stops moving, grey wolves attack it to finish the
male and a female) are the leaders of the pack and they dominate hunting process. This is modelled by decreasing a from 2 to 0 dur-
the whole pack. Other members of the pack are the followers of ing the iterations. As a is decreased, the fluctuation ranges of A
alphas. Besides the social hierarchy, hunting is also an interesting also decreases. When |A| < 1, the grey wolves attack the prey.
social behaviour of the grey wolves. The main phases of grey wolf
3.2.5. Search of prey (exploration)
hunting are: (1) Tracking, chasing, and approaching the prey (2)
Grey wolves diverge from each other in search of prey. When
Pursuing, encircling, and harassing the prey until it stops moving
a prey is found, they converge to attack it. The searching process
and (3) Attacking the prey.
is guided by the alpha, beta, and delta. When the coefficient vec-
tor |A| > 1 the grey wolves diverge (exploration) from each other
3.2. Mathematical model of GWO in search of prey. The coefficient C in Eq. (8) also favours explo-
ration by providing random weights to the prey. The natural obsta-
The mathematical model of the GWO as proposed by Mirjalili cles that make it difficult for the grey wolves to approach a prey
et al. (2014) is presented here in brief covering the social hierarchy are modelled using this parameter. It allows the GWO to explore
and hunting strategy. the solution space and helps in avoiding any local optima.
 A.K.M. Khairuzzaman, S. Chaudhury / Expert Systems With Applications 86 (2017) 64–76 67

4. Methodology

4.1. The proposed GWO based multilevel thresholding method

Grey wolves represent the search agents and their positions

represent the thresholds to be optimized. Therefore, depending
upon the number of thresholds, the wolves move in 1-D, 2-D, 3-D,
or hyper dimensional space by changing their position vectors. The
positions of the grey wolves are first initialized randomly. Then the
fitness of all the wolves is determined using Eq. (2) or Eq. (4). The
best three wolves are designated as alpha, beta, and the delta. All
other wolves update their positions determined by the positions
of the alpha, beta, and delta using Eq. (11). The positions of the
wolves are updated if better positions are found. This process is
repeated till the maximum number of iteration is completed. The
position of the alpha wolf gives the desired thresholds. The pseudo
code of the proposed multilevel thresholding method is presented
in the following subsection.

4.2. Pseudo code of the proposed GWO based multilevel thresholding

−
→
Initialize the positions of the grey wolves, X i (i = 1, 2, ..., n )
Initialize the value ofaas 2.
Calculate the coefficientsA and Cusing Eqs. (7) and (8) re-
spectively
Calculate the objective function value of each wolf by using Fig. 1. Benchmark test images used to perform experiments. These images are
Eq. (2) for Otsu, or Eq. (4) for Kapur based method taken from USC-SIPI (a–e) and BSD 500 (f–h) image databases. The images (a)–(e)
−
→ − → −
→ are of size 512 × 512 each. And the images (f)–(h) are of size 481 × 321 each. (a)
X α , X β , and X δ are the positions of alpha, beta, and delta
Lena, (b) Peppers, (c) Baboon, (d) Man, (e) Airplane, (f) Starfish, (g) Snake, and (h)
wolves
Zebra.
WHILE (t<Maximum number of iterations)
FOR each search agent Table 1
Update the position of the wolf using the Eq. (11) Parameters of the algorithms.
END FOR
Decrease linearly the values of afrom 2 to 0 during the iter- Algorithm Parameter Value
ations GWO Number of grey wolves 12
UpdateA and Cusing Eqs. (7) and (8) respectively Number of iterations 100
Calculate the objective function value of each wolf by using PSO Number of particles 12
the Eq. (2) for Otsu, or Eq. (4) for Kapur based method Number of iterations 100
−
→ − → −
→ Maximum particle velocity 6
Update X α , X β , and X δ Maximum inertia weight 0.9
t =t +1 Minimum inertia weight 0.2
END WHILE Cognitive and social parameters [0,2]
Return Xα BFO Number of bactera 12
Number of chemotactic steps 10
Number of reproductive steps 4
Number of ellimination steps 2
5. Experimental environment Probability of ellimination 0.95
Split ratio 0.5
This section presents a brief description on the experimental set Depth of attractant 0.1
Width of attractant 0.2
up of the proposed method. Firstly, a description on some bench-
Height of repellant 0.1
mark test images is presented. Parameters of different algorithms Width of repellant 10
taken into comparison with the proposed method are discussed.
Finally the segmentation validation metric is briefly described.
In this study, eight test images from different databases are that are used in this study are presented in Table 1. All the algo-
used. Lena, Peppers, Baboon, Man, and Airplane images are taken rithms are implemented in Matlab R2010a with Intel core-i7 CPU
from USC-SIPI image database which are of size 512 × 512 each. @ 3.40 GHz.
While Starfish, Snake, and Zebra images are taken from BSD 500 The quality of the segmented images are compared by using
image database and are of size 481 × 321 each. These test images Mean Structural SIMilarity index (MSSIM) (Wang, Bovik, Sheikh,
are presented in Fig. 1. Member, Simoncelli & Member, 2004). The MSSIM index evaluates
Performance of the proposed GWO based multilevel threshold- overall image quality and is calculated as,
ing method is compared with some widely used optimization al-  
gorithms such as, PSO and BFO. BFO is first proposed by Passino 
M

(2002) and PSO is first proposed by Kennedy and Eberhart (1995).

MSSIM (X, Y ) = SSIM (x j , y j ) /M (12)
j=1
The proposed method is compared with an improved version of
BFO (Sathya & Kayalvizhi, 2011b) and also with an improved ver- where X and Y represents the reference and the segmented im-
sion of PSO (Shi & Eberhart, 1998) that introduces an inertia pa- age respectively. xj and yj are the image contents at jth local win-
rameter to the original PSO. The parameters of each algorithm dow and M is the total number of local windows in the image. The
68 A.K.M. Khairuzzaman, S. Chaudhury / Expert Systems With Applications 86 (2017) 64–76

Table 2
The objective function values obtained by GWO, BFO, and PSO based methods.

Test images m Kapur’s objective values Otsu’s objective values

GWO BFO PSO GWO BFO PSO

Lena 4 18.0122 18.0082 17.9877 2191.84 2191.47 2191.84

5 20.6093 20.5975 20.6078 2217.77 2213.11 2215.99
Peppers 4 18.7338 18.7331 18.5541 3151.99 3151.84 3151.90
5 21.4016 21.3955 21.3847 3195.93 3195.11 3195.90
Baboon 4 18.1304 18.1297 18.1304 1692.14 1691.79 1692.14
5 20.7896 20.7583 20.7870 1717.89 1715.23 1717.85
Man 4 18.6810 18.6807 18.5997 3210.64 3210.57 3210.64
5 21.4749 21.4680 21.3089 3256.54 3255.96 3254.45
Airplane 4 18.3121 18.3107 18.3115 2070.07 2069.22 2070.07
5 20.9088 20.9018 20.9072 2096.13 2092.65 2096.10
Starfish 4 19.0579 19.0537 19.0563 2869.43 2869.40 2869.20
5 21.8090 21.8068 21.8072 2916.27 2914.69 2915.72
Snake 4 18.6376 18.6352 18.6356 1285.68 1285.35 1285.42
5 21.3965 21.3885 21.3835 1315.66 1315.34 1315.47
Zebra 4 17.8857 17.8805 17.8821 1590.82 1590.65 1590.74
5 20.4511 20.4491 20.4414 1618.15 1618.14 1618.12

Table 3
The threshold values obtained by GWO, BFO, and PSO based Kapur’s entropy methods.

Test images m Threshold values obtained by Kapur’s entropy methods

GWO BFO PSO

Lena 4 64,97,137,179 66,97,136,179 78,120,160,192

5 63,94,128,163,194 61,95,127,164,194 62,93,126,162,194
Peppers 4 46,84,130,179 45,84,129,178 71,108,142,180
5 43,76,111,144,181 42,75,108,142,182 45,78,115,149,184
Baboon 4 33,74,114,159 33,73,113,158 33,74,114,159
5 33,70,105,139,173 33,71,109,147,179 33,66,101,136,171
Man 4 46,90,132,175 47,90,132,175 62,117,174,230
5 46,89,131,174,230 43,87,130,174,230 46,89,131,172,206
Airplane 4 67,106,145,185 67,106,144,184 67,106,144,185
5 60,89,123,155,187 57,89,121,154,186 62,92,125,157,188
Starfish 4 68,116,164,206 69,114,163,207 66,113,160,203
5 56,93,132,170,209 56,95,135,174,212 54,92,131,170,208
Snake 4 75,123,168,211 74,122,165,211 75,123,167,211
5 68,107,146,181,219 70,110,147,183,219 62,99,138,177,217
Zebra 4 92,134,168,207 89,134,168,208 91,134,169,207
5 90,130,161,191,222 92,131,161,191,222 75,104,136,169,208

Structural Similarity index (SSIM) is given as,
SSIM (x, y ) = (2μx μy + C1 )(2σxy + C2 )/(μ + μ + C1 )
2 2
x y
× (σx2 + σy2 + C2 )
where μx and μy are the mean intensities of the image local win-
dows defined by x and y respectively. σx and σy are the standard
deviation of x and y respectively. σxy is the local correlation co-
efficient between x and y. C1 and C2 are constants. The values of
MSSIM index are in the range [−1, 1]. A higher value of MSSIM
index indicates a better quality of the segmented image.
6. Results and discussion
This section presents the experimental results of the proposed
GWO based multilevel thresholding technique and its analysis in
terms of solution quality, stability, and computational time. A sta-
tistical analysis is also performed to see the superiority of the pro-
posed GWO based technique over the other efficient algorithms. All
these are discussed in the following subsections.
6.1. Solution quality
The values of the objective function as obtained by GWO, BFO,
and PSO based methods are presented in Table 2 and the corre-
sponding threshold values are presented in Tables 3 and 4. In max-
imization algorithm, higher the value of the objective function bet-
ter is the solution. For a small number of thresholds (such as, m = 2
and 3), the values of the objective function obtained by all the al-
gorithms are practically the same. Therefore, the results with the
higher number of thresholds (such as, m = 4 and 5) are only pre-
(13) sented. It is seen from the Table 2 that GWO based methods most
often find higher value of objective function compared to BFO and
PSO based methods.
Segmented images as obtained by GWO-Kapur method and
GWO-Otsu method are presented in Figs. 2 and 3 respectively. The
segmentation results for Lena image for GWO-Kapur and GWO-
Otsu methods are presented in Figs. 4 and 5 respectively. The qual-
ity of the segmented images is then compared by using MSSIM in-
dex. The MSSIM index evaluates the visual similarity between the
original image and the segmented image. The MSSIM index values
of the segmented images obtained by all the methods are given in
the Table 5. MSSIM index gives a higher value when the segmented
image is more similar to the original image. From the Table 5 it is
found that MSSIM index values of the segmented images by GWO
based methods are higher than the BFO and PSO based methods.
For example, the MSSIM index values in case of Lena image with
five thresholds for Otsu based methods are 0.8069, 0.8035, and
0.8048 for GWO, BFO, and PSO respectively. It clearly shows that
GWO based method gives higher quality segmentation compared
to BFO and PSO based methods. It is also seen from Table 5 that,
the value of MSSIM index increases as the number of thresholds
increase. This indicates that segmentation quality improves as the
number of thresholds is increased. This is also evident from the
 A.K.M. Khairuzzaman, S. Chaudhury / Expert Systems With Applications 86 (2017) 64–76 69

Fig. 2. Segmented images obtained by GWO-Kapur multilevel thresholding method. (a)–(g) represent 3-level thresholding, (a’)–(g’) represent 4-level thresholding, (a’’)–(g’’)
represent 5-level thresholding, and (a’’’)–(g’’’) represent 6-level thresholding.
70 A.K.M. Khairuzzaman, S. Chaudhury / Expert Systems With Applications 86 (2017) 64–76

Fig. 3. Segmented images obtained by GWO-Otsu multilevel thresholding method. (a)–(g) represent 3-level thresholding, (a’)–(g’) represent 4-level thresholding, (a’’)–(g’’)
represent 5-level thresholding, and (a’’’)–(g’’’) represent 6-level thresholding.
 A.K.M. Khairuzzaman, S. Chaudhury / Expert Systems With Applications 86 (2017) 64–76 71

Fig. 4. Segmentation results for Lena image by GWO-Kapur method. (a)–(d) represent segmented images into three, four, five, and six classes respectively. (a’)–(d’) represent
the fitted histogram and the thresholds for the segmented images (a)–(d) respectively.
72 A.K.M. Khairuzzaman, S. Chaudhury / Expert Systems With Applications 86 (2017) 64–76

Fig. 5. Segmentation results for Lena image by GWO-Otsu method. (a)–(d) represent segmented images into three, four, five, and six classes respectively. (a’)–(d’) represent
the fitted histogram and the thresholds for the segmented images (a)–(d) respectively.
 A.K.M. Khairuzzaman, S. Chaudhury / Expert Systems With Applications 86 (2017) 64–76 73

Table 4
The threshold values obtained by GWO, BFO, and PSO based Otsu’s between class variance methods.

Test images m Threshold values obtained by Otsu method

GWO BFO PSO

Lena 4 75,114,145,180 74,115,145,180 75,114,145,180

5 73,109,136,160,188 67,104,126,151,186 67,97,124,151,183
Peppers 4 41,89,135,175 40,89,134,175 40,88,134,174
5 39,80,118,150,182 41,81,120,149,183 39,81,119,151,183
Baboon 4 72,106,137,168 72,104,136,167 72,106,137,168
5 67,99,125,149,174 57,88,115,141,170 67,98,124,148,174
Man 4 35,82,124,164 36,82,124,164 35,82,124,164
5 28,65,100,134,172 30,66,103,136,174 26,61,95,129,164
Airplane 4 84,129,172,203 79,125,170,202 84,129,172,203
5 69,107,143,180,205 57,104,141,182,203 70,108,144,180,205
Starfish 4 60,101,138,187 60,101,139,188 59,99,136,186
5 52,86,117,150,194 26,39,102,158,178 51,86,117,150,189
Snake 4 70,102,130,167 70,100,129,168 70,102,130,169
5 64,91,114,139,174 66,94,116,140,176 65,93,117,142,178
Zebra 4 84,112,142,199 84,111,141,199 85,113,143,200
5 79,104,127,158,209 76,101,123,150,204 75,98,121,149,203

Table 5
Comparison of MSSIM index values of the segmented images obtained by GWO, BFO, and PSO based methods. A Higher value of MSSIM
index indicates better quality of the segmented image.

Test images m MSSIM index values of Kapur methods MSSIM index values of Otsu methods

GWO BFO PSO GWO BFO PSO

Lena 4 0.7765 0.7716 0.7752 0.7814 0.7806 0.7814

5 0.7895 0.7866 0.7869 0.8069 0.8035 0.8048
Peppers 4 0.7227 0.7206 0.7214 0.7214 0.7206 0.7206
5 0.7497 0.7472 0.7482 0.7526 0.7435 0.7504
Baboon 4 0.8014 0.8012 0.8014 0.8577 0.8572 0.8577
5 0.8590 0.8556 0.8558 0.8868 0.8827 0.8852
Man 4 0.7093 0.7022 0.7082 0.7257 0.7246 0.7257
5 0.7102 0.7091 0.7089 0.7731 0.7682 0.7712
Airplane 4 0.8605 0.8596 0.8598 0.8423 0.8402 0.8423
5 0.8725 0.8702 0.8722 0.8622 0.8604 0.8610
Starfish 4 0.7455 0.7441 0.7450 0.7732 0.7721 0.7727
5 0.7957 0.7888 0.7929 0.8188 0.8177 0.8179
Snake 4 0.8138 0.8056 0.8102 0.8708 0.8700 0.8706
5 0.8496 0.8469 0.8473 0.9017 0.8987 0.9009
Zebra 4 0.7934 0.7912 0.7933 0.8393 0.8369 0.8374
5 0.8041 0.8002 0.8039 0.8625 0.8587 0.8621

Figs. 2 and 3 that the visual quality of the segmented images im-
proves as the number of thresholds increase. The thresholds are
fitted in the histogram of the segmented images in Figs. 4 and 5.
These figures illustrate how the thresholds are able to segment the
image into different classes.
As metaheuristic algorithms are stochastic in nature, the solu-
tion found at each run may not be identical. Therefore another
test is performed to analyze the accuracy and stability of the al-
gorithms using mean and standard deviation of the objective func-
tion values. A lower value of standard deviation indicates higher
stability of the algorithm. Whereas, higher mean value of the ob-
jective function indicates higher accuracy. Each algorithm is run
100 times for each 16 cases to find the standard deviation and the
mean value of the objective function. The mean values of objective
functions are given in Table 6. From the table it is observed that
GWO based method most often finds higher mean value of the
objective function than the BFO and PSO based methods. Hence,
the proposed GWO based multilevel thresholding method gives
more accurate segmentation than the BFO and PSO based meth-
ods. The standard deviation of the algorithms is given in Table 7.
It is evident from the table that GWO based method more often
has a lower standard deviation than BFO and PSO based methods.
Hence, GWO based method is more stable compared to BFO and
PSO based methods.
6.2. Statistical analysis
A statistical analysis using Wilcoxon rank sum (Wilcoxon, 1945)
test is performed at 5% significance level. The objective function
values of the proposed method are compared with BFO and PSO
based methods. All the algorithms are run 100 times for the sta-
tistical analysis. The null hypothesis is constructed as: there is no
significant difference between the three algorithms. The alternative
hypothesis considers that there is a significant difference between
the three algorithms. The p and h values are presented in Table 8.
A value of p > 0.05 (or h = 0) indicates that the null hypothesis can-
not be rejected. On the other hand, a value of p < 0.05 (or h = 1)
means the null hypothesis can be rejected at 5% significance level.
A p value where GWO performs better than the other algorithms is
marked with ‘+’ in the superscript and with ‘#’ where GWO perfor-
mance is similar or worse than the other algorithms. In the exper-
iment using Otsu function, GWO based method produces better re-
sult in 16 out of 16 cases when compared with BFO based method
and produces better result in 13 out of 16 cases when compared
with PSO based method. Whereas, in the experiment using Kapur
entropy function, GWO based method produces better result in 16
out of 16 cases when compared with BFO based method and pro-
duces better result in 15 out of 16 cases when compared with PSO
based method. These results suggest that there is a significant dif-
ference between the three algorithms. In most cases GWO based
74 A.K.M. Khairuzzaman, S. Chaudhury / Expert Systems With Applications 86 (2017) 64–76

Table 6
Mean values of the objective functions obtained by GWO, BFO, and PSO based methods at 100 run each.

Images m Mean values of Kapur’s entropy Mean values of Otsu function

GWO BFO PSO GWO BFO PSO

Lena 4 18.0013 17.9947 18.0 0 08 2191.84 2189.76 2191.84

5 20.6073 20.5675 20.6047 2217.34 2204.44 2217.21
Peppers 4 18.7284 18.7204 18.7049 3151.98 3151.63 3151.98
5 21.3881 21.3517 21.3819 3195.72 3195.05 3195.66
Baboon 4 18.1281 18.1152 18.1281 1692.14 1689.41 1692.14
5 20.7854 20.7450 20.7851 1717.81 1708.41 1717.75
Man 4 18.6738 18.6630 18.6728 3210.62 3209.95 3210.62
5 21.4388 21.3527 21.3619 3256.52 3253.68 3256.41
Airplane 4 18.3112 18.3024 18.3112 2069.94 2065.89 2069.94
5 20.9030 20.8609 20.9024 2096.12 2094.38 2096.04
Starfish 4 19.0566 19.0125 19.0421 2869.08 2867.25 2868.82
5 21.8045 21.7824 21.7836 2915.59 2915.43 2915.46
Snake 4 18.6340 18.5632 18.6291 1285.37 1285.22 1285.31
5 21.3807 21.2794 21.3472 1315.13 1315.02 1315.07
Zebra 4 17.8819 17.6284 17.7946 1590.06 1589.38 1589.76
5 20.4229 20.3971 20.4135 1617.56 1617.19 1617.42

Table 7
Standard deviation of the objective functions obtained by GWO, BFO, and PSO based methods at 100 run each.

Test images m Standard deviation of Kapur’s entropy Standard deviation of Otsu’s function

GWO BFO PSO GWO BFO PSO

Lena 4 0.0067 0.0190 0.0168 0.2101 0.1217 0.2614

5 0.0134 0.1605 0.0313 0.6718 1.0531 0.9487
Peppers 4 0.0028 0.0233 0.0910 0.0455 0.1904 0.8347
5 0.0110 0.0278 0.1178 0.0965 0.2617 1.3599
Baboon 4 0.0026 0.0136 0.0187 0.2658 1.0241 0.5534
5 0.0059 0.0158 0.04554 0.3430 1.1605 0.6295
Man 4 0.0011 0.0526 0.0406 0.0222 0.2988 0.1636
5 0.0421 0.0552 0.0811 0.0317 1.6689 1.3403
Airplane 4 0.0021 0.0264 0.0314 0.2580 1.9501 0.4872
5 0.0074 0.0274 0.0362 0.5038 2.0132 1.5311
Starfish 4 0.0018 0.0224 0.0254 0.4234 0.4351 0.6027
5 0.0028 0.0398 0.0501 0.5790 1.5473 1.6019
Snake 4 0.0031 0.0181 0.0209 0.6584 0.7323 0.2477
5 0.0138 0.0210 0.0483 0.9166 0.9924 1.1310
Zebra 4 0.0046 0.0358 0.0107 0.4894 0.7374 0.5327
5 0.0104 0.0997 0.0149 1.0176 2.1359 1.8710

Table 8
Statistical analysis (Wilcoxon Rank sum Test) of the metaheuristic algorithm based multilevel thresholding methods. 100 run for each of
the cases. p = probability of the statistic. h = 1 means the null hypothesis can be rejected at 5% level of significance. A p value where GWO
performs better than the other algorithms is marked with ‘+’ in the superscript and with ‘#’ where GWO performance is similar or worse
than the other algorithms.

Image m Otsu based method Kapur based method

GWO vs. PSO GWO vs. BFO GWO vs. PSO GWO vs. BFO

p h p h p h p h

Lena 4 >0.05# 0 <0.05+ 1 <0.05+ 1 <0.05+ 1

5 <0.05+
Peppers 4 <0.05+
5 <0.05+
Baboon 4 >0.05#
5 <0.05+
Man 4 <0.05+
5 <0.05+
Airplane 4 >0.05#
5 <0.05+
Starfish 4 <0.05+
5 <0.05+
Snake 4 <0.05+
5 <0.05+
Zebra 4 <0.05+
5 <0.05+
1 <0.05+ 1 <0.05+ 1 <0.05+ 1
1 <0.05+ 1 <0.05+ 1 <0.05+ 1
1 <0.05+ 1 <0.05+ 1 <0.05+ 1
0 <0.05+ 1 >0.05# 0 <0.05+ 1
1 <0.05+ 1 <0.05+ 1 <0.05+ 1
1 <0.05+ 1 <0.05+ 1 <0.05+ 1
1 <0.05+ 1 <0.05+ 1 <0.05+ 1
0 <0.05+ 1 <0.05+ 1 <0.05+ 1
1 <0.05+ 1 <0.05+ 1 <0.05+ 1
1 <0.05+ 1 <0.05+ 1 <0.05+ 1
1 <0.05+ 1 <0.05+ 1 <0.05+ 1
1 <0.05+ 1 <0.05+ 1 <0.05+ 1
1 <0.05+ 1 <0.05+ 1 <0.05+ 1
1 <0.05+ 1 <0.05+ 1 <0.05+ 1
1 <0.05+ 1 <0.05+ 1 <0.05+ 1
 A.K.M. Khairuzzaman, S. Chaudhury / Expert Systems With Applications 86 (2017) 64–76 75

Table 9 of the search space is high. The PSO algorithm is one of the best
Average CPU time (millisecond) of GWO, BFO, and PSO based methods at 100 run
and arguably the most widely used optimization algorithm. It has
each.
fewer parameters to adjust and it is one of the fastest optimiza-
Test images m Kapur method Otsu method tion algorithms. But it may fall into local optimum in higher di-
GWO BFO PSO GWO BFO PSO mensional search space. The GWO algorithm has few parameters
Lena 2 35.9 88.3 29.1 30.5 89.8 25.8
to adjust. The best three solutions (alpha, beta, and delta) guide
3 61.5 102.0 34.4 51.6 88.3 27.3 the searching process. Exploration and exploitation parameters are
4 100.8 103.1 46.2 77.3 162.5 29.7 perfectly balanced which enables it to avoid local optimum which
5 146.9 181.2 47.0 114.1 279.7 40.6 is the reason of the better performance of GWO over BFO and PSO.
Peppers 2 36.7 88.6 32.0 32.8 90.6 25.4
The main disadvantages (weakness) of the proposed method are
3 62.5 125.8 32.8 52.3 82.8 27.3
4 98.4 101.6 44.5 78.1 175.0 30.5 as follows. The proposed method only uses the intensity informa-
5 148.4 181.1 49.2 110.2 268.0 40.0 tion of the image to perform segmentation. It makes the method
Baboon 2 37.5 92.7 28.9 31.3 93.8 25.8 sensitive to noise and inefficient for advanced image segmenta-
3 64.8 106.3 32.8 48.4 96.9 28.1 tion problems. Thresholding based methods generally do not per-
4 103.1 107.3 46.8 77.3 184.4 32.0
form satisfactorily with images that contain intensity inhomogene-
5 151.6 174.2 49.2 114.8 293.0 41.4
Man 2 37.5 89.1 28.1 30.6 91.4 25.8 ity problem. For example, medical images that contain intensity in-
3 65.6 96.9 32.8 48.4 84.4 28.1 hogeneity, advanced segmentation techniques such as deformable
4 103.1 99.2 46.1 76.6 178.1 37.5 models (Ciecholewski, 2016; Li, Huang, Ding, Gatenby, Metaxas &
5 153.1 169.5 53.1 109.4 269.5 39.8
Gore, 2011; Li, Xu, Member, Gui & Fox, 2010; Mcinerney & Ter-
Airplane 2 35.3 83.6 29.7 32.0 93.8 26.6
3 61.7 95.3 32.8 48.4 94.4 28.1 zopoulos, 1996) are used which combines geometry, physics and
4 98.4 100.8 46.9 75.0 178.9 38.3 approximation theory to carry out segmentation.
5 151.6 186.7 49.2 107.0 288.3 41.4
Starfish 2 30.5 90.6 20.8 22.9 87.5 18.0
3 56.3 98.3 23.4 40.6 88.2 21.1 7. Conclusion
4 93.0 109.6 36.7 64.8 179.7 28.1
5 146.1 179.2 37.5 99.2 273.4 32.0
Snake 2 30.5 87.8 22.7 23.1 85.2 18.2 In this study, the merit of Grey Wolf Optimizer (GWO) in mul-
3 55.5 94.8 25.0 40.6 86.4 21.1 tilevel thresholding problem is investigated. Kapur’s entropy and
4 93.8 96.9 35.0 63.4 186.7 28.9 Otsu’s between class variance functions are considered for thresh-
5 142.2 173.4 36.6 101.6 282.8 32.0 old selection. The computational complexity of multilevel thresh-
Zebra 2 29.7 80.5 22.7 23.4 86.7 14.8
olding increases many fold with the increasing number of thresh-
3 54.7 86.3 23.4 39.8 87.3 18.2
4 90.6 89.1 34.4 64.8 170.8 19.8 olds. So, to mitigate this problem, a GWO based technique is used
5 139.1 161.4 37.3 98.4 281.3 24.2 to speed up the threshold searching process. The proposed method
is tested on some standard test images, such as, Lena, Peppers, Ba-
boon, Man, Airplane, Starfish, Snake, and Zebra. The performance
multilevel thresholding algorithm performs better than the other of the proposed method is then compared with PSO and BFO based
two algorithms. methods. The experimental results suggest that the proposed GWO
based method is more stable and yields solutions of higher quality
6.3. Computation time than BFO and PSO based methods. The computational complexity
of GWO, BFO, and PSO based methods are compared with the aver-
Average CPU time is used to compare the computational com- age CPU run time. The comparison shows that GWO is faster than
plexity of the multilevel thresholding methods. The average CPU BFO but slower than PSO.
time for GWO, BFO, and PSO based methods are given in Table 9. The Kapur method and the Otsu method are very efficient
Each of the algorithms was run 100 times to find the average CPU for bi-level thresholding, while their computational complexity in-
time using Intel core-i7 CPU @ 3.40 GHz. As for example, in case creases exponentially in case of multilevel thresholding. In order to
of Lena image with five thresholds, the average CPU time for Ka- make these methods practical for multilevel thresholding, GWO is
pur based method are 146.9, 181.2, and 47.0 ms for GWO, BFO, and used in this work. The segmentation results were analyzed using
PSO respectively. Whereas, the average CPU time for Otsu based MSSIM index, the mean, and the standard deviation of the objec-
methods are 114.1, 279.7, and 40.6 ms for GWO, BFO, and PSO re- tive functions. The experimental study reveals that the GWO algo-
spectively. From Table 9 it is seen that GWO is faster than BFO but rithm together with Kapur’s entropy or Otsus’s between class vari-
slower than PSO. ance function can be effectively used for multilevel thresholding.
The results of the proposed GWO based multilevel thresholding
6.4. Advantages and disadvantages of the proposed method method are promising. In future work, authors aim to use some
spatial properties of images along with the intensity information
The main advantages (strengths) of the proposed method are to improve thresholding based image segmentation technique. And
as follows. The proposed multilevel thresholding method is sim- also aim to finding a simpler and more efficient GWO algorithm
ple and easy to implement. Multilevel thresholding barely using and apply it to image segmentation and computer vision prob-
the Kapur’s (Kapur et al., 1985) or Otsu’s (Otsu, 1979) method are lems. Further, the merit of GWO can be investigated using mini-
computationally expensive because they search the solution space mum cross entropy, Tsallis entropy, Renyi’s entropy for multilevel
exhaustively. The proposed GWO based method greatly reduces the thresholding.
computational complexity of multilevel thresholding by optimizing
the threshold searching process. The proposed method is tested on References
intensity images. The experimental results suggest that it performs
quite well on intensity image segmentation problems. Ali, M., Ahn, C. W., & Pant, M. (2014). Multi-level image thresholding by synergetic
The BFO algorithm gives very competitive result compared to differential evolution. Applied Soft Computing Journal, 17, 1–11. https://doi.org/
10.1016/j.asoc.2013.11.018.
other algorithms. But it has many parameters to adjust and it may Alihodzic, A., & Tuba, M. (2014). Improved bat algorithm applied to multilevel image
be prone to be trapped in a local optimum when the dimension thresholding. Scientific World Journal, 2014. https://doi.org/10.1155/2014/176718.
76 A.K.M. Khairuzzaman, S. Chaudhury / Expert Systems With Applications 86 (2017) 64–76
Arora, S., Acharya, J., Verma, A., & Panigrahi, P. K. (2008). Multilevel thresholding
for image segmentation through a fast statistical recursive algorithm. Pattern
Recognition Letters, 29(2), 119–125. https://doi.org/10.1016/j.patrec.20 07.09.0 05.
Bhandari, A. K., Singh, V. K., Kumar, A., & Singh, G. K. (2014). Cuckoo search algo-
rithm and wind driven optimization based study of satellite image segmenta-
tion for multilevel thresholding using Kapur’s entropy. Expert Systems with Ap-
plications, 41(7), 3538–3560. https://doi.org/10.1016/j.eswa.2013.10.059.
Brest, J., Mlakar, U., & Poto, B. (2016). A hybrid differential evolution for optimal
multilevel image thresholding, 65, 221–232. https://doi.org/10.1016/j.eswa.2016.
08.046.
Ciecholewski, M. (2016). An edge-based active contour model using an infla-
tion/deflation force with a damping coefficient. Expert Systems With Applications,
44, 22–36. https://doi.org/10.1016/j.eswa.2015.09.013.
Cuevas, E., Sencion, F., Zaldivar, D., Perez-Cisneros, M., & Sossa, H. (2012). A multi-
threshold segmentation approach based on artificial bee colony optimization.
Applied Intelligence, 37(3), 321–336. https://doi.org/10.1007/s10489- 011- 0330- z.
Cuevas, E., Zaldivar, D., & Pérez-cisneros, M. (2010). Expert systems with applica-
tions a novel multi-threshold segmentation approach based on differential evo-
lution optimization. Expert Systems With Applications, 37(7), 5265–5271. https:
//doi.org/10.1016/j.eswa.2010.01.013.
Gao, H., Xu, W., Sun, J., & Tang, Y. (2010). Multilevel thresholding for image seg-
mentation through an improved quantum-behaved particle swarm algorithm.
IEEE Transactions on Instrumentation and Measurement, 59(4), 934–946. https:
//doi.org/10.1109/TIM.2009.2030931.
Kapur, J. N, Sahoo, P. K., & Wong, A. K. C. (1985). A new method for gray-level pic-
ture thresholding using the entropy of the histogram. Computer Vision Graphics
and Image Processing, 273–285.
Kennedy, J., & Eberhart, R. (1995). Particle swarm optimization. In Neural networks,
1995. Proceedings, IEEE international conference on: vol. 4 (pp. 1942–1948). https:
//doi.org/10.1109/ICNN.1995.488968.
Li, C., Huang, R., Ding, Z., Gatenby, J. C., Metaxas, D. N., & Gore, J. C. (2011). A
level set method for image segmentation in the presence of intensity inhomo-
geneities with application to MRI, 20(7), 2007–2016.
Li, C., Xu, C., Member, S., Gui, C., & Fox, M. D. (2010). Distance regularized level set
evolution and its application to image segmentation, 19(12), 3243–3254.
Liao, P.-S., Chen, T.-S., & Chung, P.-C. (2001). A fast algorithm for multilevel thresh-
olding. Journal of Information and Engineering, 17, 713–727.
Maitra, M., & Chatterjee, A. (2008a). A hybrid cooperative-comprehensive learning
based PSO algorithm for image segmentation using multilevel thresholding. Ex-
pert Systems with Applications, 34(2), 1341–1350. https://doi.org/10.1016/j.eswa.
20 07.01.0 02.
Maitra, M., & Chatterjee, A. (2008b). A novel technique for multilevel optimal mag-
netic resonance brain image thresholding using bacterial foraging. Measurement,
41(10), 1124–1134. https://doi.org/10.1016/j.measurement.20 08.03.0 02.
Mcinerney, T., & Terzopoulos, D. (1996). Deformable models in medical image anal-
ysis: A survey, 1(2), 91–108.
Mehmet Sezgin, B. S. (2004). Survey over image thresholding techniques and quan-
titative performance evaluation. Journal of Electronic Imaging, 13(January), 146–
165. https://doi.org/10.1117/1.1631316.
Mirjalili, S., Mohammad, S., & Lewis, A. (2014). Advances in engineering software
grey wolf optimizer. Advances in Engineering Software, 69, 46–61. https://doi.org/
10.1016/j.advengsoft.2013.12.007.
Otsu, N. (1979). A threshold selection method from gray-level histograms. IEEE
Transactions on Systems, Man and Cybernetics, 9(1), 62–66.
Pal, N. R., & Pal, S. K. (1993). A review image seg techniques. Pattern Recognition,
26(9), 1277–1294.
Passino, K. M. (2002). Biomimicry of bacterial foraging. IEEE Control System Maga-
zine, (June), 52–67.
Reddi, S. S., Rudin, S. F., & Keshavan, H. R. (1984). An optimal multiple threshold
scheme for image segmentation. IEEE Transactions on Systems, Man and Cyber-
netics, SMC-14, 4, 661–665. https://doi.org/10.1109/TSMC.1984.6313341.
Sahoo, P. K., & Soltani, W. , S. (1988). SURVEY A survey of thresholding techniques∗ .
Computer Vision, Graphics, and Image Processing, 233–260.
Sarkar, S., Das, S., & Sinha, S. (2016). Hyper-spectral image segmentation using Rényi
entropy based multi-level thresholding aided with differential evolution. Ex-
pert Systems With Applications, 50, 120–129. https://doi.org/10.1016/j.eswa.2015.
11.016.
Sathya, P. D., & Kayalvizhi, R. (2011a). Amended bacterial foraging algorithm for
multilevel thresholding of magnetic resonance brain images. Measurement,
44(10), 1828–1848. https://doi.org/10.1016/j.measurement.2011.09.005.
Sathya, P. D., & Kayalvizhi, R. (2011b). Modified bacterial foraging algorithm based
multilevel thresholding for image segmentation. Engineering Applications of Arti-
ficial Intelligence, 24(4), 595–615. https://doi.org/10.1016/j.engappai.2010.12.001.
Sathya, P. D., & Kayalvizhi, R. (2011c). Optimal multilevel thresholding using bacte-
rial foraging algorithm. Expert Systems with Applications, 38(12), 15549–15564.
https://doi.org/10.1016/j.eswa.2011.06.004.
Shannon, C. E. (1948). A mathematical theory of communication. The Bell Sys-
tem Technical Journal, 27(July), 379–423. 1928 https://doi.org/10.1145/584091.
584093 .
Shi, Y., & Eberhart, R. (1998). A modified particle swarm optimizer. In Evolution-
ary computation proceedings, 1998. IEEE world congress on computational intelli-
gence., the 1998 IEEE international conference on (pp. 69–73). https://doi.org/10.
1109/ICEC.1998.699146.
Wang, Z., Bovik, A. C., Sheikh, H. R., Member, S., Simoncelli, E. P., & Mem-
ber, S. (2004). Image quality assessment. From Error Visibility to Structural Simi-
larity, 13(4), 600–612.
Wilcoxon, F. (1945). Individual comparisons by ranking methods. International Bio-
metric Society, 1(6), 80–83.
Yin (1999). A fast scheme for optimal thresholding using genetic algorithms. Signal
Processing, 72(2), 85–95. https://doi.org/10.1016/S0165-1684(98)00167-4.
Yin, P.-Y., & Chen, L.-H. (1997). A fast iterative scheme for multilevel thresholding
methods. Signal Processing, 60, 305–313.